Emotion understanding from text has been predominantly approached through discrete classification frameworks. However, complementary perspectives in psychology and neuroscience suggest that affective states may also exhibit continuous, context-dependent, and temporally structured properties. Building on this perspective, we investigate emotion representations as latent dynamical trajectories within transformer embedding spaces. Using GoEmotions and MELD across four transformer architectures, we analyze the geometric organization and temporal evolution of learned emotional states. We find that latent trajectories exhibit temporal persistence and structured transition dynamics, while the prediction error increases during emotion shifts, acting as a signal of expectation violation. We additionally introduce trajectory-level latent interventions enabling controllable emotional transitions through directional operations in representation space. Finally, manifold capacity analysis reveals partially aligned and variably separable emotional manifolds, suggesting shared geometric organization across affective representations.